Efficient Deformable Filter Banks

Transcription

1 1168 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 46, NO. 4, APRIL 1998 Efficient Deformable Filter Banks Roberto Manduchi, Pietro Perona, and Doug Shy Abstract This correspondence describes efficient schemes for the computation of a large number of differently scaled/oriented filtered versions of an image. We generalize the well-known steerable/scalable ( deformable ) filter bank structure by imposing X-Y separability on the basis filters. The resulting systems, designed by an iterative projections technique, achieve substantial reduction of the computational cost. To reduce the memory requirement, we adopt a multirate implementation. Due to the inner sampling rate alteration, the resulting structure is not shift invariant. We introduce a design criterion for multirate deformable structures that jointly controls the approximation error and the shift variance. Index Terms Filter banks, multirate filtering, pattern analysis, steerable filters. I. INTRODUCTION Elementary visual structures such as lines, edges, texture, motion, are powerful cues to understand the structure of the outside world from its visual appearance (the image), and their identification is instrumental for almost any visual task. Classical image processing problems (enhancement, denoising) may be approached successfully using elementary descriptors such as edges and textures. Image compression schemes using sub-band coders, oriented along the preferred texture orientation in the image, proved advantageous in terms of visual rendition [1]. Velocity may be interpreted as orientation in the spatio-temporal domain, and motion-compensated spatio-temporal filters may be used successfully for prediction, interpolation, and smoothing [2], as well as for coding [3]. Regardless of the specific descriptor of interest, most techniques start processing the image (or image sequence) with a family of linear filters tuned at a wide range of orientations and scales of resolution (see [4] for an extensive bibliography on the subject). The multiscale/multiorientation image decomposition is then analyzed to detect features (usually via a nonlinear stage) and to measure their attributes (orientation, dominant scale, velocity). While orthogonal structures have been intensively studied in the context of wavelet theory, several algorithms are designed to operate on redundant (or overcomplete) image decompositions. A drawback of this approach is that the computational cost to realize the analysis filter bank may easily become too high for practical use. In order to meet prescribed implementation constraints, the use of fast filtering is mandatory. Manuscript received February 15, 1997; revised November 30, This work was supported by NSF Grant IRI on Geometry Driven Diffusions, NSF Research Initiation Grant IRI , ONR Grant N , an NSF National Young Investigator Award to P. Perona, the Center for Neuromorphic Systems Engineering as a part of the National Science Foundation Engineering Research Center Program, and the California Trade and Commerce Agency, Office of Strategic Technology. The associate editor coordinating the review of this paper and approving it for publication was Dr. Henrique Malvar. R. Manduchi is with the Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA USA. P. Perona is with the California Institute of Technology, Pasadena, CA USA and with D.E.I., Università di Padova, Padova, Italy. D. Shy is with the California Institute of Technology, Pasadena, CA USA. Publisher Item Identifier S X(98) This correspondence describes efficient schemes for the computation of a large number of differently scaled/oriented filtered versions of an image, pushing forward previous results by Perona [4] and Freeman and Adelson [5]. To reduce the computational weight of multiscale/multiorientation filter banks, one may exploit the correlation among the filters in a two-stage scheme usually named steerable or scalable [4], [5] (or more generally, deformable [4]). We point out that such systems are particular instances of the multistage separable structure introduced by Treitel and Shanks [6]; this observation leads us to consider deformable filter bank structures with X Y separable basis filters. Our implementation effectively reduces the computational weight associated with deformable filter banks. An iterative projections technique is used for the least-squares design of such structures. While deformable filter banks are efficient in terms of computational cost, they require substantial extra memory to store intermediate frames. To reduce the overall memory requirement, one may use a multirate implementation; for those filters in the filter bank that are narrow banded, only a subsampled version of the intermediate filtered frames needs to be stored. In our implementation, we embed the filter bank in an analysis/synthesis separable pyramidal scheme. Such a structure, unfortunately, is not shift invariant. Shift invariance is usually considered to be a fundamental property of signal processing systems; on the other side, a system that deviates slightly from shift invariance may still be suitable to most applications. In fact, if our task is the analysis of the visual structures in an image, aliasing may be tolerated as a signal-dependent noise, qualitatively not too different from other forms of noise that are traditionally considered. This notion is at the basis of our novel least-squares design procedure for multirate FIR filters. Such a technique relies on the definition of a suitable multirate approximation criterion, which jointly controls the approximation error and the shift variance. We extend this procedure to the design of our pyramidal deformable filter banks, achieving the result of joint reduction of computational weight and of memory requirement. The paper is organized as follows: Section II describes the leastsquares design of X Y multistage separable deformable filter banks. Section III introduces our algorithm to design least-squares multirate FIR filters and describes the design of pyramidal separable deformable filter banks. Section IV concludes the correspondence. II. SEPARABLE DEFORMABLE FILTER BANKS Consider a set of FIR filters fd(x; ); 2 S; 2 T g, composed by scaled and oriented versions of a prototype kernel. If N(; ) is the size of filter d(x; ;)(where represents the scale, and represents the orientation), in general, 2S 2T N(; ) elementary operations (sums or multiplications) per input pixel (OPP s) are required to implement the overall multiscale/multiorientation filter bank in a direct-form realization. On the other side, some classes of multidimensional FIR filters admit efficient implementation. In particular, if we are interested in analysis filter banks that represent the rotated version of a prototype kernel, we may exploit the idea of steerable filters, which was introduced by Freeman and Adelson [5] and developed further by Perona [4]. A steerable filter d(x; )is such that it can be decomposed as (see Fig. 1) d(x; )= R01 d r (x)t r (): (1) r= X/98$ IEEE

2 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 46, NO. 4, APRIL Fig. 1. Steerable filter bank. Fig. 3. Separable steerable filter bank. decomposition index (or rank) R, find the one-dimensional (1-D) kernels fu r(x); v r(y);0 r < Rg and recombination functions ft r (); 0 r<rgthat minimize the quadratic norm of the error e(x; y; ) =d(x; y; ) 0 R01 u r(x)v r(y)t r(): (3) r=0 Fig. 2. Multistage separable filter. We will call the filters d r(x) basis filters and the functions t r() recombination functions. Term R is called the rank of the decomposition. Perona [4] studied the conditions for a kernel d(x; )(with continuous x and ) to be exactly steerable. One easily shows [7] that steerable filters are formally equivalent to multistage X Y separable structures introduced by Treitel and Shanks [6] (see Fig. 2). As in the case of multistage X Y separable filters, the least-squares design of steerable filter banks [4] is performed by computing the SVD of a matrix built from the kernel d(x; ), where each row of the matrix corresponds to an orientation (similar considerations apply to the case of scalable filter banks [4]). The computational savings using the discrete steerable scheme may be computed as follows (the case of scalable filters is analogous). The direct implementation requires NN OPP s, where N is the size of the support of d(x; )(assumed constant with ), and N is the number of orientations. Using the steerable implementation, R convolutions are computed once for all; then R multiplications (and R 0 1 additions) per pixel are required for each orientation. The overall number of OPP s does not exceed R(N + N ), which is smaller than NN for sufficiently small values of R. The price to pay is that we need to store R filtered version of the image to be recombined for each orientation. A. Least-Squares Design of Separable Deformable Filter Banks The idea of multistage separability may be extended to higher dimensions, imposing X Y separability on the basis filters of the steerable/scalable filter banks. In the case of steerable filter banks, we consider decompositions of the type R01 r=0 u r(x)v r(y)t r() (2) (see Fig. 3). We will call this structure a separable steerable filter bank. The extension to scalable structures (to approximate rotated and scaled kernels) is straightforward. The approximation problem we attack can be formalized as follows: Given the set of kernels fd(x; y; ); x 2 X; y 2 Y; 2 Tg, where X; Y; T are discrete (and finite) sets, and given the Although, in the case of two-dimensional (2-D) multistage separable structures (e.g., multistage X Y separable filters or nonseparable steerable filter banks), the approximation task can be solved using the SVD of the kernels matrix, we now find ourselves dealing with more complex structures. In fact, the minimization task turns out to be a trilinear approximation problem, for which the 2-D SVD technique is not appropriate. On the other side, it is easily seen that if two variables [say, fu r (x)g and fv r (x)g] are kept fixed, then minimizing the approximation error norm is a linear problem in the third variable [here, ft r ()g]. This observation suggests the following iterative procedure: Keep two variables fixed and minimize for the third one; then, cycle for the other variables. The approximation error is guaranteed to lower at each step, and since it is bounded from below by zero, the algorithm converges to a minimum. In [7], we use the formalism of hypermatrix algebra [8] and Kronecker algebra to translate the approximation task into a sequence of simple matrix operations. Unfortunately, one cannot be sure that the minimum found this way will be global (the error surface being in general not convex), and the solution will depend on the starting point. Experimental tests and strategies for the selection of the starting point and for termination criteria are discussed in [7]. 1) Experimental Tests: We consider the separable steerable/scalable decomposition of a family of rotated and scaled kernels. The continuous prototype kernel is the separable function d(x; y) = d1(x)d2(y), where d1(x) is the second derivative of a Gaussian with standard deviation, and d2(y) is a Gaussian with standard deviation 2 = 3 (filters with similar elongate shape are widely used in computer vision to achieve high orientation selectivity). The orientation is sampled in N = 20 points for 0 < 180, whereas the scale, ranging from 1 to 4 (two octaves), is sampled logarithmically on N = 12 points. The continuous kernels are sampled on Z 2 and windowed within a square of side N = N x = N y = 35. Each scaled/oriented kernel is normalized to unitary energy. The results for R = 30 are shown in Figs Two observations may be drawn from the experimental data. 1) Our algorithm designs basis kernels all with the same size. Observing Fig. 4, however, it is clear that some basis kernels may be conveniently windowed to a smaller support without introducing relevant additional approximation error and with reduced computational cost. The design of deformable filter banks with inhomogeneous kernels sizes is described in [7]. However, note that we have the following.

3 1170 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 46, NO. 4, APRIL 1998 Fig. 4. Basis filters fu r (x)g, fv r (y)g and recombination functions fq r (; )g in the separable steerable/scalable decomposition (R =30) for the test of Section II-A1. 2) For a sufficiently high number of scales/orientations, the computational weight due to the basis filters becomes negligible with respect to that of the recombination. In fact, the ratio between the OPP s for the two operations is approximately equal to 2N=N N. III. MULTIRATE DEFORMABLE FILTER BANKS Imposing X Y separability on the basis filters of a deformable filter bank reduces the overall computational weight at the price of supplementary memory for storing the intermediate filtered images. Frame memories are expensive, and it is desirable to reduce the memory requirement while enjoying reduced computational cost. A solution is derived observing that some of the basis filters may be narrow banded. The filtered signals are thus highly correlated and may be subsampled before storing (therefore, using less memory). These reduced-rate versions are then interpolated and linearly recombined for each scale and orientation. Although this may seem computationally expensive at first, we can show that with a suitable choice of the interpolator filters, the overall implementation results are efficient both in terms of memory and of computational cost.

4 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 46, NO. 4, APRIL (a) Fig. 7. (a) Approximating and (b) original kernel for the system of Fig. 5 corresponding to the orientation that gives the highest error for =2:4 (=45 -error = 16.2%). (b) Fig. 5. Error norm per kernel versus decomposition index R for the test of Section II-A1. Dashed line: nonseparable steerable/scalable SVD-based decomposition [4]. Solid line: separable steerable/scalable decomposition with non-(-) separable recombination functions (solid line). Dotted line: separable steerable/scalable decomposition with (-) separable recombination functions. As expected, by imposing separability, we pay the price of higher approximation error for the same rank. Fig. 8. Multirate FIR filter. System of Fig. 8 with M =2redrawn using polyphase decompo- Fig. 9. sition. Fig. 10. Basic multirate structure used in the pyramidal scheme. Fig. 6. Error norm E ; = x; y je(x; y; ; )j2 versus angle and scale for the system of Fig. 5. Fig. 11. Polyphase implementation of the raised-cosine interpolator. In the case of separable deformable filter banks, we adopt the multirate implementation for those 1-D basis filters that are suitably narrow banded and jointly optimize for the multirate 1-D filters and for the recombination functions. In particular, we use the pyramidal octave-band separable structure represented in Fig. 13. A. Least Squares Design of Multirate FIR Filters The general structure of a M-fold multirate FIR filter [9] is shown in Fig. 8. Term M is the multirate order of the multirate filter. To analyze this system, the z-transform notation results are handy. We will indicate sequences in the space (time) domain with lowercase letters and their z-transform with the corresponding uppercase letters. In the case of Fig. 8, G(z) and Q(z) are two FIR filters interconnected via the cascade of an M-fold decimator followed by an M-fold interpolator. The presence of decimator/interpolator blocks in the scheme allows us to use polyphase implementation to reduce the computational weight. For example, if M = 2, we may redraw the scheme of Fig. 8 using the Type 1 polyphase decomposition [10] of the kernels involved. Here, G(z) = G 0(z 2 )+z 01 G 1(z 2 ), where g 0(x) = g(2x); g 1 (x) = g(2x +1). We thus obtain the scheme of Fig. 9, where all the filters work on reduced sampling rate, and therefore, with respect to the scheme of Fig. 8, the number of OPP s is reduced by one half. However, unless the filters are perfectly bandlimited, the system is linear periodically time variant (LPTV). Rabiner and Crochiere [9] designed multirate filters with brickwall frequency response using a minimax error criterion in the frequency

5 1172 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 46, NO. 4, APRIL 1998 This problem, unfortunately, is not in a least-squares formulation, and designing the optimal filters g(x) and q(x) for (7) involves a nonlinear procedure. We may then consider the following suboptimal task. Minimize E 2 = kp(0) (x) 0 d(x)k 2 + kp (1) (x) 0 d(x)k 2 : (8) 2 To better understand the meaning of the proposed error criterion, let us translate it in terms of t (0) (x) and t (1) (x). Using (6), we have that 2E 2 = kt (0) (x) 0 d(x)k 2 + kt (1) (x) 0 d(x)k 2 : (9) The following upper bound holds. kt (0) (x) 0 t (1) (x)k 2 2[E 2 + kt (0) (x) 0 d(x)k 1kt (1) (x) 0 d(x)k]: (10) Fig. 12. Pyramidal scheme for the multirate 1-D scalable decomposition. domain. These filters are typically used for sampling rate conversion of acoustic signals. In our case, however, we are not interested in sharp lowpass responses; therefore, the minimax criterion of [9] is not adequate for our applications. Instead, we propose a novel least-squares criterion for multirate systems based on the time domain, which explicitly takes into account the shift variance. For the following discussions, it is important to notice that the leastsquares design of deterministic linear filters in the temporal domain has the following operatorial interpretation: Approximating a target convolution kernel d(x) with a kernel u(x) in the least-squares sense is equivalent to minimizing the maximum punctual deviation from the desired output for all unit norm input signals [11]. Consider the multirate scheme of Fig. 8, where for simplicity s sake, we set M = 2. As noted earlier, the system is LPTV, and therefore, it is characterized by two impulse responses, corresponding, for instance, to inputs (x) and (x +1)[where (x) is the discrete impulse]. Let t (0) (x) and t (1) (x +1)be such impulse responses. It is easily shown that t (0) (x) =q3g 0 (x);t (1) (x +1)=q3g 1 (x) (4) where g 0 (x) and g 1 (x) are obtained by interleaving g 0 (x) and g 1 (x) with null-valued samples. If t (0) (x) =t (1) (x), the system is shift invariant, and it operates a convolution with a (fixed) kernel t (0) (x). Otherwise, inputting a signal l(x), the output at point x 0 will be y(x 0 )= hl(x);p(0) (x 0 x 0)i; even x 0 hl(x); p (1) (x 0 x 0 )i; odd x 0 (5) where h1; 1i is the inner product, and p (0) (x) = t(0) (x); even x t (1) (x); odd x p (1) (x) = t(1) (x); even x t (0) (x); odd x: Minimizing the maximum punctual deviation from the desired output, in this case, is equivalent to the following task. Minimize maxfkp (0) (x) 0 d(x)k 2 ; kp (1) (x) 0 d(x)k 2 g: (7) (6) Now, using (4), one can show that the minimization task is multilinear in g 0 (x), g 1 (x), and q(x). An iterative algorithm to minimize form (9), which is similar in spirit to the technique of Section II-A, is described in [7]. The minimization of form (8) has an interesting counterpart in the frequency domain. Using Parseval s equality, one can show that (8) is equivalent to the following task. Minimize jd(e j! ) 0 Q(e j! )G(e j! )=2j 2 d! 0 2 jq(e j! )G[e j(!+) ]=2j 2 d!: (11) The first term of (11) represents the deviation from the target response, assuming the two filters are correctly bandlimited; the second term measures the residual spectral overlapping with the repetition of G(e j! ) introduced by the inner decimation. B. Pyramidal Implementation of Deformable Filter Banks As mentioned earlier, we consider a slightly less general multirate scheme than in Fig. 8; in fact, we embed our system in a pyramidal structure. The choice of a pyramidal scheme has been inspired by the inherently self-similarity of the scaled kernels to be approximated. The basic multirate structure is represented in Fig. 10. The two systems are equivalent if we set G(z) = H(z) and Q(z) = H(z)U(z 2 ),org(z)=h(z)u(z 2 )and Q(z) =H(z). The overall pyramidal structure is shown in Fig. 12 for the 1-D scalable case and in Fig. 13 for the 2-D case. In particular, we set the decimation filter equal to the interpolator and keep it constant for all branches in the overall deformable structure. In the scheme of Fig. 12, the basis filters are characterized by two indices (i; j); index i signals the level (or depth) in the pyramid (i =0means no inner rate alteration), whereas index j enumerates the filters in the same level. A basis filter at depth i is actually a multirate filter of multirate order 2 i for which the transfer function i of the decimator/interpolator is equal to l=1 H(z2l ). Any filter suitable for pyramidal decomposition is a candidate for h(x) as long as it can be implemented with few OPP s. This is a fundamental requirement because we store the basis filters outputs in their decimated versions for parsimony of memory. For each scale and orientation, we interpolate back these signals before recombination. In our experiments, we have used a simple three-tap raisedcosine filter h(x) = [a; 1; a] for the decimator/interpolator; the

6 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 46, NO. 4, APRIL Fig. 13. Pyramidal scheme for the 2-D multirate separable scalable/steerable decomposition. optimization of h(x) is discussed in [7]. To interpolate of a factor 2 using the polyphase implementation of Fig. 11, we need only 0.5 multiplications and 0.5 sums per output sample. To appreciate the reduction of the computational burden for the recombination, consider a 1-D scalable filter bank where all R basis filters are implemented in a two-multirate scheme. For each scale/orientation, we need 0.5 (for interpolation) plus 0.5R (for recombination) multiplications per input sample. A comparison with the nonmultirate case (where the reconstruction requires R multiplications per input sample per scale/orientation) shows that even if the multirate implementation may require an increase in the filter bank s rank to compensate for larger approximation error, it makes for the substantial reduction of the computational cost associated with the basis filtering and the recombination. The problem of the a priori selection of the multirate order and kernel s size for each basis filter is discussed in [7], where we also derive an iterative technique for the least-squares design of pyramidal deformable filter banks. In brief, with respect to the case of simple separable deformable filter banks, one has to consider the effect of the decimators/interpolators and the fact that now, the basis filters are characterized by a variable number of impulse responses. Our pyramidal structure can be extended to the 2-D case for steerable/scalable decompositions. The most general scheme would consider nonseparable subsampling lattices [10]; however, we restrict ourselves to the case of separable lattices and filters (see Fig. 13). Note that if I x and I y are the maximum depths of the filters in (I +I X and in Y respectively, the system is characterized by 2 ) impulse responses. IV. CONCLUSION We have described a technique for the efficient implementation of deformable filter banks based on the use of separable basis filters and multirate implementation. The filter bank is embedded in a pyramidal structure and designed under a novel multirate error criterion that jointly minimizes the approximation error and the shift variance. The computational weight and memory requirement may be substantially reduced with respect to the common steerable/scalable decomposition based on SVD. The Matlab software to design the filter banks described in this paper may be found at manduchi/def.tar.z. REFERENCES [1] A. Ikonomopoulos and M. Kunt, High compression image coding via directional filtering, Signal Process., vol. 8, no. 2, pp , Apr [2] E. Dubois, Motion-compensated filtering of time-varying images, Multidimensional Syst. Signal Process., vol. 3, pp , [3] E. Chang and A. Zakhor, Subband video coding based on velocity filters, in Proc. IEEE Int. Symp. Circuits Syst., San Diego, CA, 1992, vol. 5, pp [4] P. Perona, Deformable kernels for early vision, IEEE Trans. Pattern Anal. Machine Intell., vol. 17, pp , May [5] W. Freeman and E. Adelson, The design and use of steerable filters, IEEE Trans. Pattern Anal. Machine Intell., vol. 13, pp , [6] S. Treitel and J. Shanks, The design of multistage separable planar filters, IEEE Trans. Geosci. Electron., vol. GE-9, pp , [7] R. Manduchi, P. Perona, and D. Shy, Efficient implementation of deformable filter banks, California Institute of Technology, 1997, Tech. Rep. CNS-TR-97-04; also available at [8] D. G. Antzoulatos and A. A. Sawchuk, Hypermatrix algebra: Theory, Comput. Vision, Graphics Image Process.: Image Understanding, vol. 57, no. 1, pp , Jan [9] L. R. Rabiner and R. E. Crochiere, A novel implementation for narrowband FIR digital filters, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-23, pp , Oct [10] P. P. Vaidyanathan, Multirate Systems and Filter Banks. Englewood Cliffs, NJ: Prentice-Hall, [11] T. W. Parks and D. P. Kolba, Interpolation minimizing maximum normalized error for band-limited signals, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-26, pp , Aug